Evaluating quantum random forest hyperparameters for high-resolution wildfire severity mapping using Sentinel-2 satellite

Our research analyzes a computational framework for evaluating hyperparameters of quantum random forests in the context of high-resolution wildfire severity mapping using Sentinel-2. Wildfires pose catastrophic ecological threats on a global scale, requiring advanced analysis techniques for the accurate assessment of environmental damage and rehabilitation planning. Traditional machine learning models often encounter limitations when dealing with complex, multidimensional ecological data at fine spatial scales. The primary objective is to compare classical hyperparameters with those optimized using quantum computing within random forest architectures in order to determine their respective effectiveness in predicting burn severity classes. The number of trees n_estimators ranges from 50 to 500, the maximum depth of individual trees max_depth ranges from 5 to 50, and the minimum samples required to split a node min_samples_split ranges from 2 to 10. The optimization procedure is implemented using quantum Scikit-learn Python libraries, designed to integrate quantum kernels and quantum-assisted optimization routines directly into standard Scikit-learn estimators. Rather than relying on classical surrogate models, our approach leverages quantum kernel estimation and variational quantum classifiers integrated into the Scikit-learn workflow. The methodology begins with the acquisition and preprocessing of Level-2A Sentinel-2 images, incorporating atmospheric and topographic corrections. Next, spectral indices such as the Normalized Burn Ratio are calculated using near-infrared and short-wave infrared bands to isolate affected vegetation. The core of the process employs quantum Scikit-learn architectures to train classification models and efficiently optimize hyperparameters. Finally, the optimized models are used to generate an innovative 3D spatial map of fire severity, enabling researchers and emergency responders to comprehensively visualize the extent of terrain-related damage. The results showed that, for the standard random forest baseline model using default hyperparameters, the classification evaluation yielded an overall accuracy of 0.812, a balanced accuracy of 0.795, and a macro F1-score of 0.801. In contrast, the quantum-optimized random forest configuration achieved an overall accuracy of 0.864, a balanced accuracy of 0.851, and a macro F1-score of 0.858, with a magnitude of improvement of 6.41% for overall accuracy, 7.04% for balanced accuracy, and 7.12% for macro F1-score.